Specify `--load-format pt` for Mixtral

[pierrestock]

As explained here, HF pushed some changes to the official repo to make it compatible with transformers==4.36.0.dev0. Need to double check that this does not break vLLM integration (in particular, we don't want to load weights twice, and model name changed from mistral to mixtral).

https://github.com/vllm-project/vllm/blob/b5f882cc98e2c9c6dde7357dbac2ec0c2c57d8cd/requirements.txt#L10C20-L10C20

[pierrestock]

Adding --load-format pt seems do to the trick. I'll let you guys (@WoosukKwon @zhuohan123) advertise it (maybe an assert in the code)?

[xiechengmude]

Adding --load-format pt seems do to the trick. I'll let you guys (@WoosukKwon @zhuohan123) advertise it (maybe an assert in the code)?

I met this problem too. How to add this param ?